For much of the past two years, the conversation around artificial intelligence has centred on speed, capability and adoption. Companies raced to introduce AI assistants, automate workplace tasks and encourage employees to use increasingly powerful models. Spending was often treated as a necessary cost of staying competitive. Now, a different question is beginning to dominate boardroom discussions: how much is all of this actually costing?
That question came into sharp focus this week when OpenAI chief executive Sam Altman acknowledged that concerns about AI spending have risen rapidly among the company’s customers. Speaking during the Intelligence at Work event, Altman said many organisations are finding their AI budgets under pressure far sooner than expected, with some clients joking that they had exhausted their planned spending for 2026 within the first quarter of the year.
The comments mark a noticeable change in tone from an industry that has largely focused on adoption rates and technical capability. Until recently, many companies appeared willing to absorb rising AI expenses on the assumption that productivity gains would eventually justify the spending. As AI use becomes more deeply embedded in daily work, however, finance teams are beginning to examine usage patterns more closely and ask whether consumption is growing faster than measurable returns.
Altman said concerns about costs were rarely raised by customers at the start of the year. According to him, that situation has changed quickly, with companies now asking OpenAI to help them obtain greater value while spending less. The remarks suggest that cost management is becoming an increasingly important part of the AI business, not only for customers but also for the companies building and selling the technology.
Rising token consumption puts pressure on corporate budgets
At the centre of the issue is the growing use of tokens, the units that AI systems process when handling prompts, generating responses and carrying out more complicated tasks. The more tokens consumed, the higher the computing demand and the larger the bill.
The scale of that growth has been striking. Altman said that six and a half years ago OpenAI’s largest token user consumed around 100,000 tokens per month, a figure that was considered exceptional at the time. Today, he said, that volume roughly matches average per-person usage across the world. OpenAI’s largest internal user now consumes about 100 billion tokens each month, while an external user reportedly exceeds even that amount.
Such numbers illustrate how rapidly AI use has expanded inside organisations. What began as occasional experimentation has become a routine part of software development, customer service, content production, research and workplace administration. As more employees gain access to AI systems, token consumption can rise dramatically without companies fully appreciating the financial impact until invoices arrive.
Several examples have highlighted the scale of spending. Peter Steinberger, creator of OpenClaw, revealed that his team spent approximately $1.3 million on OpenAI API tokens within a single month, consuming more than 600 billion tokens during that period. Reports have also suggested that some OpenAI employees themselves consume hundreds of billions of tokens over short periods while testing and developing products.
At many companies, enthusiasm for AI has created incentives that may unintentionally increase usage. Internal competitions, performance targets and productivity initiatives have sometimes encouraged employees to route more tasks through AI systems, even when the benefits are unclear. The result has been rising consumption without a direct link to improved business outcomes.
Recent reports suggest that some firms are beginning to reassess those practices. Amazon employees reportedly acknowledged using AI tools for unnecessary tasks in order to improve their standing on internal leaderboards. Microsoft has reportedly reduced access to certain coding assistants because of growing expenses. Other technology companies have begun reviewing how AI resources are allocated across teams.
The issue is particularly important because many advanced AI applications require repeated interactions with models. Agent-based systems, which can perform sequences of tasks rather than respond to a single prompt, often consume far more tokens than traditional chatbots. Each action, verification step and follow-up request adds to overall usage, increasing computing demand and operating costs.
Search for efficiency becomes the next battleground
The growing focus on costs arrives at a time when AI companies are investing heavily in larger models, new products and expanded computing capacity. Those investments require substantial spending on data centres, specialised chips and electricity. AI providers therefore face pressure from both sides: customers want lower bills while developers continue to face expensive training and inference costs.
Altman indicated that OpenAI is working to improve efficiency and reduce costs for customers. While he did not provide specific details, the company has previously introduced smaller models, pricing adjustments and technical improvements intended to lower computing requirements.
The challenge facing the industry is that lower costs do not always lead to lower spending. Economists often point to Jevons paradox, the idea that when a resource becomes cheaper, people tend to use more of it. AI appears to be following a similar pattern. As models become less expensive to operate, organisations often expand their usage rather than reduce expenditure.
This pattern can already be seen in software development. Coding assistants have become commonplace among engineering teams, helping programmers generate code, review changes and troubleshoot problems. Each interaction may save time, but multiplied across thousands of employees and millions of requests, total consumption can rise sharply.
Questions are also emerging about how companies measure the return on their AI investments. Some executives continue to argue that higher productivity will eventually justify spending levels. Others have become more cautious, noting that increased usage does not automatically translate into better products, stronger sales or higher profits.
Uber chief executive Dara Khosrowshahi recently suggested there is not yet a clear relationship between heavy AI spending and successful product delivery. Similar concerns have surfaced across industries as companies attempt to determine where AI creates measurable value and where it simply adds cost.
